[CodeGen][ARM64EC] Add support for hybrid_patchable attribute.

llvm/include/llvm/Bitcode/LLVMBitCodes.h

@@ -757,6 +757,7 @@ enum AttributeKindCodes {
   ATTR_KIND_RANGE = 92,
   ATTR_KIND_SANITIZE_NUMERICAL_STABILITY = 93,
   ATTR_KIND_INITIALIZES = 94,
+  ATTR_KIND_HYBRID_PATCHABLE = 95,
 };
 
 enum ComdatSelectionKindCodes {

llvm/include/llvm/CodeGen/AsmPrinter.h

@@ -892,14 +892,14 @@ class AsmPrinter : public MachineFunctionPass {
   virtual void emitModuleCommandLines(Module &M);
 
   GCMetadataPrinter *getOrCreateGCPrinter(GCStrategy &S);
-  virtual void emitGlobalAlias(const Module &M, const GlobalAlias &GA);
   void emitGlobalIFunc(Module &M, const GlobalIFunc &GI);
 
 private:
   /// This method decides whether the specified basic block requires a label.
   bool shouldEmitLabelForBasicBlock(const MachineBasicBlock &MBB) const;
 
 protected:
+  virtual void emitGlobalAlias(const Module &M, const GlobalAlias &GA);
   virtual bool shouldEmitWeakSwiftAsyncExtendedFramePointerFlags() const {
     return false;
   }

llvm/include/llvm/IR/Attributes.td

@@ -112,6 +112,9 @@ def ElementType : TypeAttr<"elementtype", [ParamAttr]>;
 /// symbol.
 def FnRetThunkExtern : EnumAttr<"fn_ret_thunk_extern", [FnAttr]>;
 
+/// Function has a hybrid patchable thunk.
+def HybridPatchable : EnumAttr<"hybrid_patchable", [FnAttr]>;
+
 /// Pass structure in an alloca.
 def InAlloca : TypeAttr<"inalloca", [ParamAttr]>;
 

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

@@ -727,6 +727,8 @@ static uint64_t getAttrKindEncoding(Attribute::AttrKind Kind) {
     return bitc::ATTR_KIND_HOT;
   case Attribute::ElementType:
     return bitc::ATTR_KIND_ELEMENTTYPE;
+  case Attribute::HybridPatchable:
+    return bitc::ATTR_KIND_HYBRID_PATCHABLE;
   case Attribute::InlineHint:
     return bitc::ATTR_KIND_INLINE_HINT;
   case Attribute::InReg:

llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp

@@ -2859,8 +2859,8 @@ bool AsmPrinter::emitSpecialLLVMGlobal(const GlobalVariable *GV) {
     auto *Arr = cast<ConstantArray>(GV->getInitializer());
     for (auto &U : Arr->operands()) {
       auto *C = cast<Constant>(U);
-      auto *Src = cast<Function>(C->getOperand(0)->stripPointerCasts());
-      auto *Dst = cast<Function>(C->getOperand(1)->stripPointerCasts());
+      auto *Src = cast<GlobalValue>(C->getOperand(0)->stripPointerCasts());
+      auto *Dst = cast<GlobalValue>(C->getOperand(1)->stripPointerCasts());
       int Kind = cast<ConstantInt>(C->getOperand(2))->getZExtValue();
 
       if (Src->hasDLLImportStorageClass()) {

llvm/lib/Target/AArch64/AArch64Arm64ECCallLowering.cpp

@@ -21,6 +21,7 @@
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/IR/CallingConv.h"
+#include "llvm/IR/GlobalAlias.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Instruction.h"
 #include "llvm/IR/Mangler.h"
@@ -70,15 +71,21 @@ class AArch64Arm64ECCallLowering : public ModulePass {
   Function *buildEntryThunk(Function *F);
   void lowerCall(CallBase *CB);
   Function *buildGuestExitThunk(Function *F);
-  bool processFunction(Function &F, SetVector<Function *> &DirectCalledFns);
+  Function *buildPatchableThunk(GlobalAlias *UnmangledAlias,
+                                GlobalAlias *MangledAlias);
+  bool processFunction(Function &F, SetVector<GlobalValue *> &DirectCalledFns,
+                       DenseMap<GlobalAlias *, GlobalAlias *> &FnsMap);
   bool runOnModule(Module &M) override;
 
 private:
   int cfguard_module_flag = 0;
   FunctionType *GuardFnType = nullptr;
   PointerType *GuardFnPtrType = nullptr;
+  FunctionType *DispatchFnType = nullptr;
+  PointerType *DispatchFnPtrType = nullptr;
   Constant *GuardFnCFGlobal = nullptr;
   Constant *GuardFnGlobal = nullptr;
+  Constant *DispatchFnGlobal = nullptr;
   Module *M = nullptr;
 
   Type *PtrTy;
@@ -672,6 +679,66 @@ Function *AArch64Arm64ECCallLowering::buildGuestExitThunk(Function *F) {
   return GuestExit;
 }
 
+Function *
+AArch64Arm64ECCallLowering::buildPatchableThunk(GlobalAlias *UnmangledAlias,
+                                                GlobalAlias *MangledAlias) {
+  llvm::raw_null_ostream NullThunkName;
+  FunctionType *Arm64Ty, *X64Ty;
+  Function *F = cast<Function>(MangledAlias->getAliasee());
+  SmallVector<ThunkArgTranslation> ArgTranslations;
+  getThunkType(F->getFunctionType(), F->getAttributes(),
+               Arm64ECThunkType::GuestExit, NullThunkName, Arm64Ty, X64Ty,
+               ArgTranslations);
+  std::string ThunkName(MangledAlias->getName());
+  if (ThunkName[0] == '?' && ThunkName.find("@") != std::string::npos) {
+    ThunkName.insert(ThunkName.find("@"), "$hybpatch_thunk");
+  } else {
+    ThunkName.append("$hybpatch_thunk");
+  }
+
+  Function *GuestExit =
+      Function::Create(Arm64Ty, GlobalValue::WeakODRLinkage, 0, ThunkName, M);
+  GuestExit->setComdat(M->getOrInsertComdat(ThunkName));
+  GuestExit->setSection(".wowthk$aa");
+  BasicBlock *BB = BasicBlock::Create(M->getContext(), "", GuestExit);
+  IRBuilder<> B(BB);
+
+  // Load the global symbol as a pointer to the check function.
+  LoadInst *DispatchLoad = B.CreateLoad(DispatchFnPtrType, DispatchFnGlobal);
+
+  // Create new dispatch call instruction.
+  Function *ExitThunk =
+      buildExitThunk(F->getFunctionType(), F->getAttributes());
+  CallInst *Dispatch =
+      B.CreateCall(DispatchFnType, DispatchLoad,
+                   {UnmangledAlias, ExitThunk, UnmangledAlias->getAliasee()});
+
+  // Ensure that the first arguments are passed in the correct registers.
+  Dispatch->setCallingConv(CallingConv::CFGuard_Check);
+
+  Value *DispatchRetVal = B.CreateBitCast(Dispatch, PtrTy);
+  SmallVector<Value *> Args;
+  for (Argument &Arg : GuestExit->args())
+    Args.push_back(&Arg);
+  CallInst *Call = B.CreateCall(Arm64Ty, DispatchRetVal, Args);
+  Call->setTailCallKind(llvm::CallInst::TCK_MustTail);
+
+  if (Call->getType()->isVoidTy())
+    B.CreateRetVoid();
+  else
+    B.CreateRet(Call);
+
+  auto SRetAttr = F->getAttributes().getParamAttr(0, Attribute::StructRet);
+  auto InRegAttr = F->getAttributes().getParamAttr(0, Attribute::InReg);
+  if (SRetAttr.isValid() && !InRegAttr.isValid()) {
+    GuestExit->addParamAttr(0, SRetAttr);
+    Call->addParamAttr(0, SRetAttr);
+  }
+
+  MangledAlias->setAliasee(GuestExit);
+  return GuestExit;
+}
+
 // Lower an indirect call with inline code.
 void AArch64Arm64ECCallLowering::lowerCall(CallBase *CB) {
   assert(Triple(CB->getModule()->getTargetTriple()).isOSWindows() &&
@@ -727,17 +794,57 @@ bool AArch64Arm64ECCallLowering::runOnModule(Module &Mod) {
 
   GuardFnType = FunctionType::get(PtrTy, {PtrTy, PtrTy}, false);
   GuardFnPtrType = PointerType::get(GuardFnType, 0);
+  DispatchFnType = FunctionType::get(PtrTy, {PtrTy, PtrTy, PtrTy}, false);
+  DispatchFnPtrType = PointerType::get(DispatchFnType, 0);
   GuardFnCFGlobal =
       M->getOrInsertGlobal("__os_arm64x_check_icall_cfg", GuardFnPtrType);
   GuardFnGlobal =
       M->getOrInsertGlobal("__os_arm64x_check_icall", GuardFnPtrType);
+  DispatchFnGlobal =
+      M->getOrInsertGlobal("__os_arm64x_dispatch_call", DispatchFnPtrType);
+
+  DenseMap<GlobalAlias *, GlobalAlias *> FnsMap;
+  SetVector<GlobalAlias *> PatchableFns;
 
-  SetVector<Function *> DirectCalledFns;
+  for (Function &F : Mod) {
+    if (!F.hasFnAttribute(Attribute::HybridPatchable) || F.isDeclaration() ||
+        F.hasLocalLinkage() || F.getName().ends_with("$hp_target"))
+      continue;
+
+    // Rename hybrid patchable functions and change callers to use a global
+    // alias instead.
+    if (std::optional<std::string> MangledName =
+            getArm64ECMangledFunctionName(F.getName().str())) {
+      std::string OrigName(F.getName());
+      F.setName(MangledName.value() + "$hp_target");
+
+      // The unmangled symbol is a weak alias to an undefined symbol with the
+      // "EXP+" prefix. This undefined symbol is resolved by the linker by
+      // creating an x86 thunk that jumps back to the actual EC target. Since we
+      // can't represent that in IR, we create an alias to the target instead.
+      // The "EXP+" symbol is set as metadata, which is then used by
+      // emitGlobalAlias to emit the right alias.
+      auto *A =
+          GlobalAlias::create(GlobalValue::LinkOnceODRLinkage, OrigName, &F);
+      F.replaceAllUsesWith(A);
+      F.setMetadata("arm64ec_exp_name",
+                    MDNode::get(M->getContext(),
+                                MDString::get(M->getContext(),
+                                              "EXP+" + MangledName.value())));
+      A->setAliasee(&F);
+
+      FnsMap[A] = GlobalAlias::create(GlobalValue::LinkOnceODRLinkage,
+                                      MangledName.value(), &F);
+      PatchableFns.insert(A);
+    }
+  }
+
+  SetVector<GlobalValue *> DirectCalledFns;
   for (Function &F : Mod)
     if (!F.isDeclaration() &&
         F.getCallingConv() != CallingConv::ARM64EC_Thunk_Native &&
         F.getCallingConv() != CallingConv::ARM64EC_Thunk_X64)
-      processFunction(F, DirectCalledFns);
+      processFunction(F, DirectCalledFns, FnsMap);
 
   struct ThunkInfo {
     Constant *Src;
@@ -755,14 +862,20 @@ bool AArch64Arm64ECCallLowering::runOnModule(Module &Mod) {
           {&F, buildEntryThunk(&F), Arm64ECThunkType::Entry});
     }
   }
-  for (Function *F : DirectCalledFns) {
+  for (GlobalValue *O : DirectCalledFns) {
+    auto GA = dyn_cast<GlobalAlias>(O);
+    auto F = dyn_cast<Function>(GA ? GA->getAliasee() : O);
     ThunkMapping.push_back(
-        {F, buildExitThunk(F->getFunctionType(), F->getAttributes()),
+        {O, buildExitThunk(F->getFunctionType(), F->getAttributes()),
          Arm64ECThunkType::Exit});
-    if (!F->hasDLLImportStorageClass())
+    if (!GA && !F->hasDLLImportStorageClass())
       ThunkMapping.push_back(
           {buildGuestExitThunk(F), F, Arm64ECThunkType::GuestExit});
   }
+  for (GlobalAlias *A : PatchableFns) {
+    Function *Thunk = buildPatchableThunk(A, FnsMap[A]);
+    ThunkMapping.push_back({Thunk, A, Arm64ECThunkType::GuestExit});
+  }
 
   if (!ThunkMapping.empty()) {
     SmallVector<Constant *> ThunkMappingArrayElems;
@@ -785,7 +898,8 @@ bool AArch64Arm64ECCallLowering::runOnModule(Module &Mod) {
 }
 
 bool AArch64Arm64ECCallLowering::processFunction(
-    Function &F, SetVector<Function *> &DirectCalledFns) {
+    Function &F, SetVector<GlobalValue *> &DirectCalledFns,
+    DenseMap<GlobalAlias *, GlobalAlias *> &FnsMap) {
   SmallVector<CallBase *, 8> IndirectCalls;
 
   // For ARM64EC targets, a function definition's name is mangled differently
@@ -837,6 +951,16 @@ bool AArch64Arm64ECCallLowering::processFunction(
         continue;
       }
 
+      // Use mangled global alias for direct calls to patchable functions.
+      if (GlobalAlias *A = dyn_cast<GlobalAlias>(CB->getCalledOperand())) {
+        auto I = FnsMap.find(A);
+        if (I != FnsMap.end()) {
+          CB->setCalledOperand(I->second);
+          DirectCalledFns.insert(I->first);
+          continue;
+        }
+      }
+
       IndirectCalls.push_back(CB);
       ++Arm64ECCallsLowered;
     }

llvm/lib/Target/AArch64/AArch64AsmPrinter.cpp

@@ -201,6 +201,7 @@ class AArch64AsmPrinter : public AsmPrinter {
   void PrintDebugValueComment(const MachineInstr *MI, raw_ostream &OS);
 
   void emitFunctionBodyEnd() override;
+  void emitGlobalAlias(const Module &M, const GlobalAlias &GA) override;
 
   MCSymbol *GetCPISymbol(unsigned CPID) const override;
   void emitEndOfAsmFile(Module &M) override;
@@ -1263,6 +1264,32 @@ void AArch64AsmPrinter::emitFunctionEntryLabel() {
   }
 }
 
+void AArch64AsmPrinter::emitGlobalAlias(const Module &M,
+                                        const GlobalAlias &GA) {
+  if (auto F = dyn_cast_or_null<Function>(GA.getAliasee())) {
+    // Global aliases must point to a definition, but unmangled patchable
+    // symbols are special and need to point to an undefined symbol with "EXP+"
+    // prefix. Such undefined symbol is resolved by the linker by creating
+    // x86 thunk that jumps back to the actual EC target.
+    if (MDNode *Node = F->getMetadata("arm64ec_exp_name")) {
+      StringRef ExpStr = cast<MDString>(Node->getOperand(0))->getString();
+      MCSymbol *ExpSym = MMI->getContext().getOrCreateSymbol(ExpStr);
+      MCSymbol *Sym = MMI->getContext().getOrCreateSymbol(GA.getName());
+      OutStreamer->beginCOFFSymbolDef(Sym);
+      OutStreamer->emitCOFFSymbolStorageClass(COFF::IMAGE_SYM_CLASS_EXTERNAL);
+      OutStreamer->emitCOFFSymbolType(COFF::IMAGE_SYM_DTYPE_FUNCTION
+                                      << COFF::SCT_COMPLEX_TYPE_SHIFT);
+      OutStreamer->endCOFFSymbolDef();
+      OutStreamer->emitSymbolAttribute(Sym, MCSA_Weak);
+      OutStreamer->emitAssignment(
+          Sym, MCSymbolRefExpr::create(ExpSym, MCSymbolRefExpr::VK_None,
+                                       MMI->getContext()));
+      return;
+    }
+  }
+  AsmPrinter::emitGlobalAlias(M, GA);
+}
+
 /// Small jump tables contain an unsigned byte or half, representing the offset
 /// from the lowest-addressed possible destination to the desired basic
 /// block. Since all instructions are 4-byte aligned, this is further compressed

llvm/lib/Target/AArch64/AArch64CallingConvention.td

@@ -333,7 +333,7 @@ def CC_AArch64_Win64_CFGuard_Check : CallingConv<[
 
 let Entry = 1 in
 def CC_AArch64_Arm64EC_CFGuard_Check : CallingConv<[
-  CCIfType<[i64], CCAssignToReg<[X11, X10]>>
+  CCIfType<[i64], CCAssignToReg<[X11, X10, X9]>>
 ]>;
 
 let Entry = 1 in

llvm/lib/Transforms/Utils/CodeExtractor.cpp

@@ -932,6 +932,7 @@ Function *CodeExtractor::constructFunction(const ValueSet &inputs,
       case Attribute::DisableSanitizerInstrumentation:
       case Attribute::FnRetThunkExtern:
       case Attribute::Hot:
+      case Attribute::HybridPatchable:
       case Attribute::NoRecurse:
       case Attribute::InlineHint:
       case Attribute::MinSize: